entrep

This puts the human even more out of the loop I'll guess?

Benchmarks saturate faster than they can be constructed by humans. Isn't that a sign of an impending singularity?

I just turned my PTZ security camera into a VAR booth for my kid. Lawn (soccer) / Driveway (Floorball) / End. Live view, step back 10s, slowmo, frame step, scrubbing.

GLM 5.3 helped to root my camera so I could get proper endpoints for PTZ and stream. Claude built the api and frontend, hosted on Home Assistant.

Coding agent that seems to beat Claude Code on SWE-bench at half the cost.

8/15 on SWE-bench Verified vs Claude Code's 5/15, ~$0.06/instance vs ~$0.12. Small sample, single repo, lots of caveats. But the direction feels right. Event-sourced reducer, no framework deps beyond the Anthropic SDK.

One could try:

> export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50

Which will have Claude Code auto compact at ~500k window size.

One could argue that subscription based inference might differ from per-token billed API usage.

Not really.

> Z-mode means the front and rear wings are closed which generates more downforce for the corners. In X-mode, the drivers can open the flaps which will reduce drag and increase speed.

Clearly, it just leaked the election results from the wrong branch of the wavefunction.

Great, thanks. This is already in my minimum requirements.

Basically I want to build it with focus on speed and work efficiency from start. To not bias myself too much, I will refrain from doing too much market research. First of all I'm building this for myself, and I'm guessing it might translate into at least a tiny market share.